The Marginal Mean of a Zero-Inflated Count

Why the Count-Component Mean Is Not the Observed Mean

A zero-inflated count model describes two ways an observed count can arise. An observation may belong to a structural-zero class, in which case its count is zero; otherwise, its count follows a specified count distribution. Consequently, the mean of that count distribution is not generally the mean of the observed counts. Some observations are set to zero before the count component can contribute anything.

The distinction matters when reporting expected counts in context. A count-component mean might describe the expected number of events among observations in the count-generating class. The marginal mean describes the expected number across all observations at the specified covariate values, including structural zeros. This tutorial derives that marginal mean and shows how to calculate it from fitted model values.

For a zero-inflated Poisson model, the count-generating distribution is Poisson. One common predictor specification is

Here \(x_i\) and \(z_i\) are the covariate vectors for the count and zero-inflation components, respectively; \(\beta\) and \(\gamma\) are their coefficient vectors. The predictors need not use the same covariates. The marginal-mean calculation below applies to any count-generating distribution with finite mean \(\mu_i\), not only the Poisson distribution. In the examples, fitted values are evaluated at specified covariates; no averaging over random effects is involved.

Deriving the Marginal Mean

The defining mixture assumptions are that a structural-zero observation has count zero, and that the count-generating class has conditional mean \(\mu_i\). The probability of the latter class is \(1-\pi_i\). Apply the law of total expectation, conditioning on class membership:

This argument does not require the count-generating distribution to be Poisson: it uses only its mean. It does require the mixture interpretation stated above. In particular, \(\pi_i\) is the probability of structural-zero class membership, not the probability that the observed count equals zero. A count-generating observation can itself have a sampling zero.

For the log-link and logit-link predictors above, the response-scale calculation is

The predictor values must first be converted to probabilities and means on their respective response scales. Multiplying a linear predictor by another fitted quantity, or multiplying the count mean by \(\hat{\pi}_i\), does not give the marginal mean.

Calculating from Linear Predictors

Model software may provide fitted linear predictors rather than fitted component means and probabilities. For a log link, exponentiating the count predictor gives \(\hat{\mu}_i\). For a logit link, applying the inverse-logit function gives \(\hat{\pi}_i\). The marginal mean is then calculated from those two response-scale values.

Comparing Predictions Across Covariate Patterns

The marginal mean combines information from both model components. A difference in predicted marginal means can reflect a change in the count-component mean, a change in the structural-zero probability, or changes in both. It is therefore important to calculate the product for each covariate pattern before comparing predictions.

When predictions are needed for multiple observations, calculate the marginal mean separately at each observation’s covariates. If the target is the expected total across \(n\) observations, add their fitted marginal means. If the target is the average expected count across those observations, average the fitted marginal means.

Distinctions, Diagnostics, and Limitations

Three quantities are easy to confuse: the count-component mean \(\mu_i\), the marginal mean \((1-\pi_i)\mu_i\), and the probability of a positive observed count. In a zero-inflated Poisson model, the last quantity is

For example, if \(\pi_i=0.25\) and \(\mu_i=2\), the marginal mean is \(0.75(2)=1.50\) counts, while the probability of a positive count is \(0.75(1-\exp(-2))\approx0.6485\). A mean is measured in counts; a probability is dimensionless. They answer different questions.

A useful prediction check is to compare fitted marginal means with observed average counts across groups defined by relevant covariate patterns. Also check the model’s predicted zero frequencies against observed zero frequencies: agreement in the overall mean alone does not ensure that zeros are well represented. For a zero-inflated Poisson model, the model-implied zero probability is

because an observed zero can arise from either class. Persistent discrepancies in means or zero frequencies can indicate that the specified predictors or count distribution do not adequately describe the data. Convergence problems or fitted probabilities near zero or one also call for careful review; they do not by themselves establish that the mixture interpretation is appropriate.

An ordinary Poisson GLM specifies a single Poisson mean rather than this two-class mixture. It can provide a different distributional description even when its predicted mean is similar. Conversely, matching the observed average with a zero-inflated model does not demonstrate that the data contain a distinct structural-zero class. The marginal-mean formula is a consequence of the assumed mixture, not a test that validates the assumption.

The calculation gives a fitted mean at specified covariates under the stated mixture model. It is not an individual-level prediction that the next observation will equal that mean, and it does not quantify uncertainty in the estimated mean. In models with additional sources of variation, predictions may require further averaging; the product derived here applies at the level where \(\pi_i\) and \(\mu_i\) are specified.